[earslap]

It is more obvious when taken to extreme: With the current feedforward transformer architectures, there is a fixed amount of compute per token. Imagine asking a very hard question with a yes/no answer to an LLM. There are infinite number of cases where the compute available to the calculation of the next token is not enough to definitively solve that problem, even given "perfect" training.

You can increase the compute for allowing more tokens for it to use as a "scratch pad" so the total compute available will be num_tokens * ops_per_token but there still are infinite amount of problems you can ask that will not be computable within that constraint.

But, you can offload computation by asking for the description of the computation, instead of asking for the LLM to compute it. I'm no mathematician but I would not be surprised to learn that the above limit applies here as well in some sense (maybe there are solutions to problems that can't be represented in a reasonable number of symbols given our constraints - Kolmogorov Complexity and all that), but still for most practical (and beyond) purposes this is a huge improvement and should be enough for most things we care about. Just letting the system describe the computation steps to solve a problem and executing that computation separately offline (then feeding it back if necessary) is a necessary component if we want to do more useful things.